Thinking in Space: How Multimodal Large Language Models See, Remember, and Recall Spaces